CMakeLists.txt 10.5 KB
Newer Older
1
2
3
4
5
# ckProfiler
set(PROFILER_SOURCES
    profiler.cpp
    profile_gemm.cpp
    profile_reduce.cpp
6
    profile_groupnorm_bwd_data.cpp
rocking's avatar
rocking committed
7
    profile_groupnorm_fwd.cpp
8
    profile_layernorm_bwd_data.cpp
9
10
    profile_layernorm_bwd_gamma_beta.cpp
    profile_groupnorm_bwd_gamma_beta.cpp
rocking's avatar
rocking committed
11
    profile_layernorm_fwd.cpp
jakpiase's avatar
jakpiase committed
12
    profile_max_pool2d_fwd.cpp
13
    profile_pool3d_fwd.cpp
14
15
    profile_avg_pool3d_bwd.cpp
    profile_max_pool3d_bwd.cpp
16
17
    profile_avg_pool2d_bwd.cpp
    profile_max_pool2d_bwd.cpp
18
19
20
    profile_softmax.cpp
    profile_batchnorm_fwd.cpp
    profile_batchnorm_bwd.cpp
21
    profile_batchnorm_infer.cpp
22
    profile_conv_tensor_rearrange.cpp
arai713's avatar
arai713 committed
23
    profile_transpose.cpp
24
    profile_permute_scale.cpp
25
)
26

27
if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
  if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
    list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp)
    list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp)
  endif()
  if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
    list(APPEND PROFILER_SOURCES profile_gemm_reduce.cpp)
    list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp)
    list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp)
    list(APPEND PROFILER_SOURCES profile_gemm_add.cpp)
    list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp)
    list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp)
    list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp)
    list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp)
    list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp)
    list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp)
    list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp)
    list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
    list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp)
46
    list(APPEND PROFILER_SOURCES profile_grouped_gemm_two_stage.cpp)
47
    list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
48
    list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp)
49
    list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp)
50
51
  endif()
  list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp)
52
  if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
53
54
55
    list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp)
    list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp)
  endif()
56
57
58
59
60
  list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp)
  list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp)
  list(APPEND PROFILER_SOURCES profile_gemm_add_multiply.cpp)
  list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp)
  list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp)
61
  list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp)
62
  list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp)
63
  list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp)
64
65
66
67
  list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp)
  list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu_add.cpp)
  list(APPEND PROFILER_SOURCES profile_conv_bwd_data.cpp)
  list(APPEND PROFILER_SOURCES profile_conv_fwd.cpp)
68
  list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd_outelementop.cpp)
69

70
endif()
71

72
if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9")
73
74
75
76
77
78
  if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
    list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp)
  endif()
  list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd.cpp)
  list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_data.cpp)
  list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp)
79
endif()
80

81
82
83
if(DL_KERNELS)
  list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp)
  list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp)
84
85
endif()

86
87
88
89
set(PROFILER_EXECUTABLE ckProfiler)

add_executable(${PROFILER_EXECUTABLE} ${PROFILER_SOURCES})
target_compile_options(${PROFILER_EXECUTABLE} PRIVATE -Wno-global-constructors)
90
91
92
93
94
# flags to compress the library
if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132)
  message("Adding --offload-compress flag for ${PROFILER_EXECUTABLE}")
  target_compile_options(${PROFILER_EXECUTABLE} PRIVATE --offload-compress)
endif()
95

96
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt)
97
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance)
rocking's avatar
rocking committed
98
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance)
99
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance)
100
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance)
101
102
103
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
jakpiase's avatar
jakpiase committed
104
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool2d_fwd_instance)
rocking's avatar
rocking committed
105
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance)
106
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool2d_bwd_instance)
107
108
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance)
109
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance)
110
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance)
arai713's avatar
arai713 committed
111
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance)
112
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance)
113

114
if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
  if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
  endif()
  if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance)
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance)
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance)
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance)
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance)
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance)
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance)
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance)
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance)
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance)
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
133
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_tile_loop_instance)
134
135
136
137
  endif()
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
138
  if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
139
140
141
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance)
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance)
  endif()
142
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
143
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance)
144
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance)
145
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance)
146
147
148
149
150
151
152
153
154
155
156
157
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
158
159
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convscale_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance)
160
161
endif()

162
if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12")
163
164
165
166
167
168
169
170
171
  if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
  endif()
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
endif()
172

173
174
if(DL_KERNELS)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance)
175
176
177
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
178
endif()
179

180
rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)